Two Wrongs Don’t Make a Right: Combating Confirmation Bias in Learning with Label Noise
نویسندگان
چکیده
Noisy labels damage the performance of deep networks. For robust learning, a prominent two-stage pipeline alternates between eliminating possible incorrect and semi-supervised training. However, discarding part noisy could result in loss information, especially when corruption has dependency on data, e.g., class-dependent or instance-dependent. Moreover, from training dynamics representative method DivideMix, we identify domination confirmation bias: pseudo-labels fail to correct considerable amount labels, consequently, errors accumulate. To sufficiently exploit information mitigate wrong corrections, propose Robust Label Refurbishment (Robust LR)—a new hybrid that integrates pseudo-labeling confidence estimation techniques refurbish labels. We show our successfully alleviates both label noise bias. As result, it achieves state-of-the-art across datasets types, namely CIFAR under different levels synthetic mini-WebVision ANIMAL-10N with real-world noise.
منابع مشابه
Two Wrongs Do Not Make a Right
491 National Tax Journal Vol. LIX, No. 3 September 2006 Abstract This paper analyzes proposals to remedy tax–induced distortions in health care by using new tax incentives and retaining all of the existing distortionary tax incentives. In the process of remedying some distortions, this approach magnifi es others—most notably increasing the total tax preference for health care. The paper conside...
متن کاملWith phases: how two wrongs can sometimes make a right
In isolation, both weak isomorphous/anomalous difference signals from heavy-atom derivatization and phases from partial molecular-replacement solutions for a subset of the asymmetric unit often fall short of producing interpretable electron-density maps. Phases generated from very partial molecular-replacement models (if generated carefully) can be used to reliably locate heavy-atom sites, even...
متن کاملذخیره در منابع من
با ذخیره ی این منبع در منابع من، دسترسی به آن را برای استفاده های بعدی آسان تر کنید
ژورنال
عنوان ژورنال: Proceedings of the ... AAAI Conference on Artificial Intelligence
سال: 2023
ISSN: ['2159-5399', '2374-3468']
DOI: https://doi.org/10.1609/aaai.v37i12.26725